In the new, connected economy, it has become increasingly important for companies or service providers to become more in tune with their clients and customers. Such contact can be facilitated with automated telephonic transaction systems, in which interactively-generated prompts are played in the context of a telephone transaction, and the replies of a human user are recognized by an automatic speech recognition system. The answers given by the respondent are processed by the system in order to convert the spoken words to meaning, which can then be utilized interactively, or stored in a database.
In order for a computer system to recognize the words that are spoken and convert these words to text, the system must be programmed to phonetically break down the words and convert portions of the words to their textural equivalents. Such a conversion requires an understanding of the components of speech and the formation of the spoken word. The production of speech generates a complex series of rapidly changing acoustic pressure waveforms. These waveforms comprise the basic building blocks of speech, known as phonemes. Vowel and consonant sounds are made up of phonemes and have many different characteristics, depending on which components of human speech are used. The position of a phoneme in a word has a significant effect on the ultimate sound generated. A spoken word can have several meanings, depending on how it is said. Speech scientists have identified allophones as acoustic variants of phonemes and use them to more explicitly define how a particular word is formed.
While there are several distinct methods for analyzing the spoken word and extracting the information necessary to enable the recognition system to convert the speech to word-strings, including Hidden Markov modeling and neural networks, these methods generally perform similar operations. The differences in these methods are typically in the manner in which the system determines how to break the phonetic signal into portions that define phonemes. Generally, a speech recognition system first converts an incoming analog voice signal into a digital signal. The second step is called feature extraction, wherein the system analyzes the digital signal to identify the acoustic properties of the digitized signal. Feature extraction generally breaks the voice down into its individual sound components. Conventional techniques for performing feature extraction include subband coding Fast Fourier Transforms and Linear Predictive Coding. Once the signal has been analyzed, the system then determines where distinct acoustic regions occur. The goal of this step is to divide the acoustic signal into regions that will be identified as phonemes which can be converted to a textual format. In isolated word systems, this process is simplified, because there is a pause after each word. In continuous speech systems, however, this process is much more difficult, since there typically are no breaks between words in the acoustic stream. Accordingly, the system must be able not only to break the words themselves into distinct acoustic regions, but must also be able to separate consecutive words in the stream. It is in this step that conventional methods such as Hidden Markov modeling and neural networks are used. The final step involves comparing a specific acoustic region, as determined in the previous step, to a known set of templates in a database in order to determine the word or word portion represented by the acoustic signal region. If a match is found, the resulting textual word is output from the system. If one is not, the signal can either be dynamically manipulated in order to increase the chances of finding a match, or the data can be discarded and the system prompted to repeat the query to the respondent, if the associated answer cannot be determined due to the loss of the data.
Many factors can adversely affect the accuracy of a speech recognition system, including the use of slang, regional dialects and accents, speech impediments, poor telephone connections and background noise. In many cases, even the most advanced speech recognition system will not be able to process such speech, and thus the loss of data is inevitable. One approach to this problem has been to limit the functionality of the speech recognition application. In this case, the users of the system must be trained to accommodate themselves to the limitations imposed by the system (e.g., state only a limited set of word replies). An extreme example of such a limited system is one in which the only response recognized by the system is a simple word or key stroke on the telephone, for example the prompt “Press or say one.” At the other extreme is a system in which a human monitors the automatic speech recognition transaction for correctness. However, such a system defeats the entire purpose of an automated system, which is generally intended to reduce costs.